Expand description
Codebook/grid tables for the IQ1_S / IQ1_M / IQ2_XXS / IQ2_XS / IQ2_S / IQ3_XXS / IQ3_S quantization formats.
These numeric tables are part of the formats themselves – a file
quantized with them cannot be decoded without these exact values, and
they cannot be independently re-derived. They are reproduced from the
MIT-licensed ggml project’s ggml-common.h (the format’s defining
source; see docs/THIRD_PARTY_NOTICES.md). The code that uses them is
written independently in lib.rs against ggml’s published dequant
semantics.
Generated from the upstream format tables – do not hand-edit.
Constants§
- IQ1S_
GRID - 2048 8-value rows, each u64 packing 8 signed int8 values.
- IQ2S_
GRID - 1024 8-value rows, each u64 packing 8 unsigned byte magnitudes. Indexed by IQ2_S’s 10-bit grid code.
- IQ2XS_
GRID - 512 8-value rows, each u64 packing 8 unsigned byte magnitudes. Indexed by IQ2_XS’s 9-bit grid code.
- IQ2XXS_
GRID - 256 8-value rows, each u64 packing 8 unsigned byte magnitudes.
- IQ3S_
GRID - 512 4-value rows, each u32 packing 4 unsigned byte magnitudes. Indexed by IQ3_S’s 9-bit grid code.
- IQ3XXS_
GRID - 256 4-value rows, each u32 packing 4 unsigned byte magnitudes.
- KMASK_
IQ2XS - Per-element sign-mask bits used by every IQ2/IQ3 sign decoding path.
- KSIGNS_
IQ2XS - 7-bit sign patterns (bit 7 = parity) shared by IQ2_XXS/IQ2_XS/IQ3_XXS. IQ2_S and IQ3_S skip this table: they store a full 8-bit sign byte per 8 elements instead of a 7-bit index into it.